AI may matter most when bioweapons researchers already have expertise

The risk depends on who is asking

For an inexperienced person or a terrorist group working in a makeshift lab, chatbots could help gather knowledge, plan experiments and troubleshoot. The most plausible work would involve existing pathogens and toxins.

The more immediate concern, says Filippa Lentzos, a science and international security specialist at King’s College London, is not AI replacing expertise but amplifying it. Well-equipped organisations and states already have trained scientists, secure laboratories and the necessary materials. AI tools could help researchers identify genetic changes that might make existing pathogens more deadly.

A more distant risk is that highly capable groups could use AI to develop entirely new pathogens or toxins that resist drugs or have no antidote.

Evidence of attempts, not proof of capability

Last year, Indian police uncovered an alleged bioterror plot. Investigators said a doctor and several others used chatbot prompts and AI-powered search while trying to make ricin, a deadly toxin, with kitchen utensils. It is not known whether the AI responses helped. Other reports, however, suggest that these tools can provide useful and dangerous information.

In September, Anthropic said it had identified five cases in which scientists from unnamed countries tried to use Claude for potentially dangerous work involving viruses or toxins. The company pointed to attempts to get around the chatbot’s safeguards after it refused requests.

In one case, Anthropic said, researchers used Claude to plan experiments that would introduce mutations into a highly pathogenic bird flu virus, with the aim of making it better adapted to mammals and easier to spread. Such work can point to efforts to make pathogens more dangerous, but similar research also takes place in legitimate laboratories to anticipate pandemic threats and develop vaccines and medicines.

Other studies have tested what chatbots might disclose. In a recent MIT experiment, students without scientific training were asked whether they could get a chatbot to help unleash a pandemic. Within an hour, chatbots had named four potential pandemic pathogens, explained how to make them from synthetic DNA and identified DNA synthesis companies that might not screen such orders. Carrying out the work in practice would be much harder.

Stanford researcher David Relman said an AI system had explained how to alter a pathogen to make it resistant to treatment. Another scientist said a chatbot had described how to use a weather balloon to drop biological weapons from the air.

The gap between an answer and a weapon

Biology research has dual uses, making it hard to distinguish work that could improve health from efforts to develop weapons. In 2025, scientists at Los Alamos National Laboratory trained an AI system to predict how potent cone-snail neurotoxins bind to human cells, work that could help create antidotes. The US government raised concerns after Chinese researchers used the work to develop their own toxin-design tool. The Chinese scientists said their goal was drug development.

In April, scientists created the first AI-designed viruses. They infect only bacteria, and the research could lead to new medicines. Experts nevertheless warned that it raised “urgent questions of biosafety and biosecurity.”

What these examples do not show is that AI can remove the practical barriers to making a weapon. Lentzos points to tacit knowledge, access to biological materials and equipment, experimental validation, scaling and delivery. Sonia Ben Ouagrham-Gormley, deputy director of the biodefense program at George Mason University in Virginia, says tacit knowledge — practical experience that is not written down and is unlikely to be available to AI — played an important role in successful state biological-weapons programmes in the past.

University of Zurich professor Martin Pacesa makes a similar point about materials: anyone with a physics degree can sketch a more efficient nuclear bomb, but without uranium the plan goes nowhere. The same applies to biological weapons. If someone already has the materials and equipment, AI may not change the situation much.

I think the sharper concern is therefore not a chatbot turning a novice into a bioweapons maker. It is whether AI lowers the cost of work for people who already have expertise and access — while making it harder to tell where legitimate research ends.

Safeguards still rely on gaps

Some AI companies have taken steps to prevent their tools from being used to develop biological weapons. Most, however, do not say whether they are doing anything, says Tom Inglesby, director of the Johns Hopkins Center for Health Security. No country has an official government policy aimed at preventing this kind of misuse.

Inglesby supports independent official assessments of new AI models before release, with identified risks addressed before public access. He also argues that access to high-risk data — such as information about how pathogens could spread faster — should be restricted.

Companies that make genetic material for researchers should screen orders to check that they are not being used to create new, especially dangerous viruses. Some screening already exists, but it is largely voluntary. AI could also help researchers create harmful genetic sequences that companies do not screen for.

That leaves a policy problem with no single checkpoint: model testing, access to sensitive data and DNA-order screening all matter, but each can leave gaps the others do not close.